Prior-Guided Multi-Scale Temporal Modeling for Behavior-Driven Residential Load Forecasting

Accurate residential load forecasting is crucial for enhancing the efficiency and reliability of energy systems in smart grid and demand response applications. However, residential load data are characterized by strong stochasticity, high volatility, and pronounced multi-scale temporal dynamics while being highly susceptible to noise and outliers. These challenges hinder existing methods from effectively capturing complex temporal patterns and learning reliable inter-variable dependencies, thereby limiting forecasting accuracy and stability. To address these issues, this paper proposes a Prior-Guided Multi-Scale Neural Network (PG-MSNN) for multi-step residential load forecasting. The proposed framework integrates prior-guided dependency modeling with multi-scale temporal representation learning in an end-to-end trainable architecture. Specifically, a learnable periodic prior space is constructed, within which a Prior-Guided Module (PGM) is designed to learn cross-variable dependencies and provide structured global periodic guidance. In parallel, a Multi-Scale Patch-LSTM Encoder (MS-PLE) is developed to model temporal dynamics across multiple scales through patch-based sequence representation and adaptive cross-scale fusion. Extensive experiments on three real-world datasets, including IHEPC, REC, and CN-OBEE, demonstrate that, under within-household temporal forecasting settings, the proposed method achieves consistent and competitive performance across various forecasting horizons.